SenseNova-U1.5
SenseTime's 8B native unified model: it reads images, draws up to 4K, edits photos and mixes text with pictures. Apache 2.0.
In short. SenseNova-U1.5 is an open model from SenseTime that both understands pictures and makes them. It can draw from a text prompt, edit a photo, answer questions about an image and write text and pictures in one reply. It has no separate image encoder, it draws at up to 4K, and its 35 GB of weights run on one 24 GB graphics card if you let it offload layers to system memory.
Summary
SenseTime's SenseNova team published SenseNova-U1.5-8B-MoT under Apache 2.0 on 20 August 2026, and the Hugging Face repository first appeared a day earlier. It is the second generation of the SenseNova-U series, which began with U1 in April 2026, and it is built on NEO-unify, an architecture that drops both the vision encoder and the VAE that most image models use. A native mixture of transformers handles understanding and generation in one model, trained with next-token prediction for text and pixel flow matching for images. U1.5 replaces the independent patch output of U1 with a convolutional reconstruction step that reduces seams at high resolution, and it generates at up to 4K. It creates images from text, edits them, takes bounding boxes and visual markers for region-level control, handles several reference images, writes interleaved text and images, and answers questions about pictures. The language side is initialized from Qwen3 weights.
SenseNova's paper reports these results from its own tables, and some rival scores are SenseNova's reproductions. On GenEval the model scores 0.92 overall, against 0.91 for U1, 0.90 for HiDream-O1-Image, 0.87 for the 20B Qwen-Image and 0.89 for the closed GPT-Image-2. On LongText-Bench it scores 0.988 in English and 0.989 in Chinese, against 0.979 and 0.962 for U1 and 0.976 and 0.928 for the 32B Emu3.5. It averages 0.948 on CVTG-2K, which the paper calls the best of the models it evaluated. On GenEval2 it scores 0.55, or 0.71 with prompt enhancement, behind the closed GPT-Image-2 at 0.81 and Qwen-Image-3.0 at 0.76. For understanding, the 8B thinking variant scores 73.86 on MMMU, 85.85 on MathVista and 90.47 on MMBench, above Qwen3-VL 8B on MathVista and MMBench, slightly below it on MMMU, and behind Qwen3.5 9B on several tests. The model card lists its remaining weaknesses: over-emphasized detail or color, errors in dense or mixed Chinese and English text, imperfect exact counts and alignment, unstable small faces and hands, and drift in long multi-turn or multi-reference edits.
The checkpoint is 35.1 GB in BF16, and Hugging Face counts 17.5B parameters, although SenseNova calls it an 8B model. The reference code needs Python 3.11, PyTorch 2.8 and CUDA 12.8, and the model loads with custom code. The README's fast memory mode streams layers from system memory so that a 24 GB card runs at near-full speed, and other modes trade speed for less VRAM. The README lists no GGUF for the final release, only a community Q8 build of 19.9 GB for the earlier Preview checkpoint, and an official 8-step LoRA speeds up generation. SenseNova's own serving stack, built on LightLLM and LightX2V, reports about 9 seconds for a 2048 by 2048 image on H100 or H200 GPUs in a single node with TP2 and CFG2 parallelism. A TailSteer version, published on 4 November 2026, is a further training stage that its authors say reduces drift across repeated edits, with 6.99 dB higher PSNR outside the edited area at the tenth turn on their own benchmark. In the 30 days to 11 November 2026, Hugging Face counted about 5,200 downloads of the main repository, so adoption is still small.
Good for
- Posters, infographics and designs with bilingual English and Chinese text, where its text-rendering scores are strongest
- Image editing with several reference images, bounding boxes or visual markers
- Pipelines that need both image generation and image understanding from a single open model
- Commercial work, since Apache 2.0 has no use restrictions of its own
Watch out for
- All scores are SenseNova's own, and on GenEval2 the model trails the closed GPT-Image-2 and Qwen-Image-3.0
- The 24 GB path relies on layer offloading and needs enough system memory, and by our estimate the whole model on the GPU needs about 40 GB
- The reference code is custom and needs a specific Python, PyTorch and CUDA setup
- Small faces, hands, dense small text and long chains of edits are still unreliable, according to the model card
- The only GGUF the README lists for U1.5 is a community build of the earlier Preview, not the official release
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of SenseNova-U1.5. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| SenseNova-U1.5-8B-MoT BF16 | 17.5B | 35 GB | 24 GB | Gaming GPU |
| SenseNova-U1.5-8B-MoT-TailSteer BF16 | 17.5B | 35 GB | 24 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 24 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is SenseNova-U1.5?
- SenseNova-U1.5 is an open model from SenseTime that both understands pictures and makes them. It can draw from a text prompt, edit a photo, answer questions about an image and write text and pictures in one reply. It has no separate image encoder, it draws at up to 4K, and its 35 GB of weights run on one 24 GB graphics card if you let it offload layers to system memory.
- Can I use SenseNova-U1.5 commercially?
- Yes. SenseNova-U1.5 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is SenseNova-U1.5 open source?
- SenseNova-U1.5 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run SenseNova-U1.5 on my own computer?
- Yes, on a good graphics card. The smallest variant, SenseNova-U1.5-8B-MoT, needs about 24 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was SenseNova-U1.5 released?
- SenseNova-U1.5 was released by SenseTime on August 20, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- SenseNova-U1.5-8B-MoT model card (Hugging Face) huggingface.co
- SenseNova-U1.5: Towards Native Unified Visual Intelligence (arXiv 2609.11929) arxiv.org
- SenseNova-U1 repository: README, inference and memory modes (GitHub) github.com
- NEO-unify: Building Native Multimodal Unified Models End to End (Hugging Face blog) huggingface.co
- SenseNova-U1.5-8B-MoT-TailSteer model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen-Image-2.1 Alibaba Qwen | 7.1B | – | Qwen Research License (Non-commercial) | 20 GB Gaming GPU | |
| HiDream-O1-Image HiDream.ai | 8.8B | – | MIT (Permissive) | 20 GB Gaming GPU | |
| Ming-Image 0.1 Design inclusionAI (Ant Group) | 6.2B | – | MIT (Permissive) | 80 GB Workstation | |
| MOSS-Image OpenMOSS | 5.08B | – | Apache 2.0 (Permissive) | 4.2 GB Laptop or phone |